Federated Learning for Lidar Super Resolution on Automotive Scenes

Abstract

In this paper, the problem of lidar super-resolution is explored under a federated learning perspective. The high cost of high-resolution lidar sensors is a major obstacle to the widespread adoption of connected and autonomous vehicles (CAVs). To reduce the cost, this study investigates the use of low-cost sensors in conjunction with super-resolution algorithms. Unlike previous studies that approach this problem with centralized solutions, this work employs federated learning to leverage private lidar data from different autonomous vehicles under different environmental conditions, resulting in more robust and diverse deep learning model. Extensive experiments on a real-world lidar odometry dataset highlight the merits and applicability of the proposed lidar super-resolution method with federated learning.